MiniMax H3 API Pricing: $0.13/s at 2K, and Why Drafting Cheap Saves Nothing

MiniMax H3 bills $0.13 per second of 2K output and $0.08 at 768P. The regeneration tier that turns a finished 768P clip into 2K costs $0.05. Add the last two and you get $0.13 — exactly the direct 2K rate, to the cent. That arithmetic is the most useful thing to know about this model’s pricing, and it is not on the marketing page.

Model ID:      MiniMax-H3
Output:        $0.13/s at 2K, $0.08/s at 768P
Regeneration:  $0.05/s (768P → 2K only)
Duration:      4–15 seconds, integers only, 24 fps
Audio:         native stereo, generated in the same pass
Input audio:   free
Input images:  first 5 free, then $0.04 each
Input video:   billed by its own duration at the output rate
Endpoint:      POST /v2/video_generation (async, poll for result)
Weights:       H3-Base open (768P); 2K module and Context-IR not released
Gateways:      not in the Ofox catalog as of 2026-09-06

Rates read from MiniMax’s pay-as-you-go pricing page on 6 September 2026. H3 launched 31 July 2026 and the weights followed on 3 August.

What MiniMax H3 Actually Is

A multimodal generation model that takes text, images, video and audio in one request and returns video with sound. Not a language model — the M3 that shows up in the same search results is a separate coding and language model that launched at the same WAIC 2026 event.

The thing that distinguishes H3 from most of the video field is that audio is generated in the same pass, not dubbed on afterwards. MiniMax’s own framing is that it collapsed what used to be separate expert models — text-to-video, image-to-video, first-and-last-frame, subject reference, motion reference, video editing — into one model where the mode follows from what you send rather than from a different endpoint.

The Pricing Table

Two output tiers, one regeneration tier, and input material billed separately.

There is also a separate MiniMax-H3-Max model at $0.05/s for 480P and $0.08/s for 768P, which currently supports text-to-video and image-to-video only and does not bill input materials.

What a clip costs at the headline rates:

The Arithmetic Nobody Puts on the Pricing Page

$0.08 + $0.05 = $0.13. Drafting at 768P and upgrading the keeper to 2K costs exactly what shooting 2K directly costs.

This matters because “generate cheap, upgrade the winners” is the obvious workflow, and it sounds like a saving. On a per-clip basis it is not one:

Where the draft path does save money: on takes you throw away. Every discarded second costs $0.05 instead of $0.13, so the saving is real but it is a function of your reject rate, not of the workflow itself. Bin four of five 5-second takes and you save $1.00 on that batch. Keep them all and you paid the same and waited through two passes instead of one.

Where it costs more: reference-heavy jobs. Regeneration rebills the original input materials. Images get five free again and then $0.025 each, and reference video is billed again at $0.05 per second of its own duration. If you are feeding a character sheet and a motion reference into every generation, the draft-then-upgrade path is strictly more expensive than going straight to 2K.

The honest rule: use 768P as a selection layer, not as a cost optimisation. It pays when you are exploring and expect to discard most of what you generate. It costs you when you already know what you want.

The Regeneration Endpoint Is Strict

Worth reading before you build around it, because it is not an upscaler. From MiniMax’s own API reference, the source video must meet the H3 768P output specification exactly:

Two more constraints that will bite:

• You must resubmit the exact original inputs. All reference images, videos and audio from the 768P job have to be passed again, plus one video_url item with role=base_video.
• The prompt must be the post-processing prompt. MiniMax is explicit that text must be “the final prompt actually sent to the model when generating the 768P source video, not the original prompt from before H3-Context-IR processing.” If you did not capture that intermediate prompt, you cannot reproduce the job.

That last one is the trap. Context-IR rewrites your instruction before generation, and regeneration wants the rewritten version. Log it at generation time or you lose the ability to upgrade the clip later.

The Open Weights Are Half the Model

H3-Base is public. The two modules that produce the headline specs are not.

MiniMax open-sourced H3 on 3 August 2026 — two BF16 checkpoints (FL2VA for text and first/last-frame work, Ref2VA for reference work), a 33B dense single-stream transformer, on Hugging Face under MiniMax’s own Community License rather than Apache or MIT.

What you do not get:

• H3-Regenerate-2K is “not yet open-sourced” per the model card. This is the module that produces 2K. A local install generates at 768P.
• H3-Context-IR, the hosted layer that turns your multimodal instruction into what the model consumes, is excluded because it spans several hosted services. MiniMax calls it critical to output quality.
• Sparse-attention inference. The model supports it natively but the open release ships full attention only.

So self-hosting gets you the 768P half. The official reference deployment is four GPUs via SGLang. Before committing an engineer to it, weigh that against $0.13 per second — and note the Community License restricts use in several territories and requires written authorisation above a revenue threshold, which is a legal review rather than a technical one.

How H3 Compares on Price

H3 is not in the Ofox catalog, so this is a cross-route comparison rather than a like-for-like one. Ofox rates were read from the live /v1/models endpoint on 6 September 2026; H3’s are MiniMax’s own list price.

Read that table carefully, because the resolution column is doing work. H3 is the only model here that outputs 2K, and it is also the only one whose cheapest tier tops out below 1080p. Comparing $0.13 for 2K against $0.05 for Wan 3.0 at 1080p is not comparing the same product.

Where each one lands:

• You need 2K with synced audio: H3 is the option on this list. Nothing else here goes above 1080p except Seedance 2.0, which reaches 4K at $0.07.
• You need 1080p at the lowest rate: Wan 3.0 at $0.05 is less than half H3’s 768P tier and outputs a higher resolution than it.
• You need clips longer than 15 seconds: H3 caps at 15. Wan 3.0 and Seedance 2.5 both go to 30.
• You need heavy reference material: Seedance 2.5 accepts 30 images, 10 videos and 10 audio clips against H3’s 9 / 3 / 3.

Calling It

H3 runs on MiniMax’s own platform, asynchronously — create a task, poll for the result:

curl -X POST https://api.minimax.io/v2/video_generation \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "content": [
      {"type": "text", "text": "A paper boat drifting down a rain gutter, close on the water line"}
    ],
    "resolution": "2K",
    "duration": 5,
    "ratio": "16:9"
  }'

Two practical notes: for pure text-to-video, ratio is required and cannot be adaptive. And keep the request body under 64 MB — Base64 inflates payloads by roughly a third, so use public URLs for anything substantial.

Through Ofox, H3 is not available as of 6 September 2026. The MiniMax models in the catalog are the M-series language models, including minimax/minimax-m3 at $0.60 / $2.40 per million tokens — the model H3 keeps getting confused with. For video on the same key today:

curl -X POST https://api.ofox.io/v1/videos \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "alibaba/wan-3.0",
    "prompt": "A paper boat drifting down a rain gutter, close on the water line",
    "duration": 5,
    "resolution": "1080p"
  }'

GET https://api.ofox.io/v1/models is the authority on what is callable, rather than this page.

What to Watch

Three things would change the read above:

1. H3-Regenerate-2K opening. Until it does, self-hosting means 768P, and the “open weights” story is materially smaller than the headline suggests. MiniMax has given no date.
2. Independent quality measurement. Everything published about H3’s output quality so far is either MiniMax’s own demos or hands-on impressions. There is no third-party video benchmark here comparable to what Artificial Analysis does for language models, so “less than a third of mainstream models” is vendor framing rather than a verified claim.
3. Gateway availability. H3 on an aggregator would let you A/B it against Wan 3.0 and Seedance 2.5 on one key, which is the comparison this page cannot run today.

Sources

• https://www.minimax.io/blog/minimax-h3
• https://platform.minimax.io/docs/guides/pricing-paygo
• https://platform.minimax.io/docs/api-reference/video-generation-v2-create
• https://platform.minimax.io/docs/api-reference/video-generation-v2-regeneration
• https://huggingface.co/MiniMaxAI/MiniMax-H3
• https://ofox.io/models/alibaba/wan-3.0
• https://ofox.io/models/bytedance/seedance-2.5

H3 rates, duration bounds, input-material rules and the regeneration source-video specification were read from MiniMax’s pay-as-you-go pricing page and API reference on 6 September 2026. Launch date 31 July 2026 and open-weights date 3 August 2026 are from MiniMax’s own posts. Ofox comparison rates and video attributes were read from the live /v1/models endpoint the same day. MiniMax H3 is not in the Ofox catalog as of that date; nothing on this page is our own benchmark run.